DeepSeek Accelerates the GenAI Industry into a New Phase: New Compute Power Growth Approach and Application Diversification

Despite the diminishing returns from scaling laws, AI compute power can continue to grow through optimized training approaches, facilitating the exploration of new applications.

A marked drop in API/token prices benefits AI software/services and on-device AI, thereby speeding up the diversification of AI applications.

For the past two years, AI server supply chain stock investments have been driven by expectations of continued growth based on scaling laws. Now, as the limitations of these scaling laws become more evident, the market is turning its attention to DeepSeek's optimized training approach - achieving major performance gains without relying on scaling laws.

According to the widely cited Chinchilla Scaling Law, AI model performance is driven by three key factors: model parameter (N), training data (D), and computing power (C). Ideal results come from scaling all three factors together.

The diminishing returns of scaling laws can be attributed to several key factors. The supply of human-created text data (D) is nearly exhausted, while simply increasing model parameters (N) proves ineffective without matching improvements in computing power (C) and addressing data volume constraints (D). Additionally, substantial short-term increases in computing power (C) face major hurdles, from delays in Blackwell system production to power supply limitations.

What makes DeepSeek-R1 particularly noteworthy from an industry research perspective is its ability to achieve significant performance improvements through optimized training approaches rather than scaling laws. These improvements can be verified through examination and testing of its open-source releases.

As the diminishing returns of scaling laws become increasingly evident, improving model performance through optimized training approaches has become essential. This shift enables sustainable growth in AI computing infrastructure and unlocks the potential for new applications. These complementary developments are crucial for the long-term advancement of the AI industry.

Currently, profits in the GenAI sector primarily stem from infrastructure provision (‘selling shovels’) and cost reduction rather than from creating new business models or generating significant value-add to existing products and services.

DeepSeek-R1 has adopted an aggressive pricing strategy. It is free to use, and the lowest API/token pricing is less than 1/100 of OpenAI-o1. This competitive pressure is likely to drive down AI usage costs. Given the intense competition in China's AI market, other Chinese firms are expected to launch LLMs with impressive benchmark scores and even more aggressive pricing.

The recent sharp correction in AI supply chain stocks primarily reflects investors' reset expectations for AI server shipments - a response to the diminishing returns of scaling laws rather than concerns over LLM service providers' and CSPs' path to profitability. Investors largely remain patient regarding profit materialization.

Declining costs in AI software/services and on-device AI, driven by lower API/token prices and optimized training approaches, are stimulating AI compute demand while alleviating investor concerns about AI investment profitability.

While declining AI prices will undoubtedly drive higher usage, it remains unclear whether volume growth can offset price reductions. Increased AI adoption may enable profitable business models, though success is not guaranteed. These uncertainties currently appear manageable as investors maintain their patience regarding profitability.

Scaling laws are empirical observations. Moderating expectations and taking a rational view of them can actually benefit long-term investment trends. Improvements such as chip upgrades (C), resolving power supply constraints (C), and the use of multimodal training data (D) in training all have the potential to re-accelerate the benefits of scaling laws in the future.

The diminishing returns from scaling laws only affect those operating at a massive scale, proving Nvidia's leadership position. When Nvidia's solutions speed up returns of scaling laws again, the company's edge over competitors like ASIC and AMD will likely become even more pronounced.

Recent production issues with the GB200 NVL72 make this a good time to reset expectations for scaling laws and AI server shipments. This stock correction will be more favorable for positive future reactions to the GB300/Rubin series.

Leading CSPs won't reduce capital expenditure simply because of improved training approaches, as efficiency gains and infrastructure investments can coexist. Any reduction in capital spending now could leave them vulnerable to falling behind competitors when scaling law benefits accelerate again.

The combined effect of open-source resources and the highly competitive environment in China is expected to lead to other Chinese companies launching high-performing LLMs with even more aggressive pricing. If LLM service providers have not established stable profitability by then, their profit pressures will intensify.

Benefiting from significant API/token price declines, AI software/service and on-device AI will attract greater investor attention. Whether this becomes a new long-term investment trend depends on whether profitable business models can be created.

Nvidia remains the likely winner when the benefits of scaling laws accelerate again in the future. However, monitoring short-term GB200 NVL72 production issues and potential changes to US semiconductor export restrictions in the medium-to-long term is crucial.